ID-insensitive deepfake detection model based on multi-attention mechanism.

Area of Science:

• Computer Science
• Artificial Intelligence
• Digital Forensics

Background:

• Deepfake technology facilitates the online spread of manipulated facial content, posing significant societal risks.
• Current deepfake detection methods often neglect the interplay between local and global image features and fail to mitigate identity leakage, leading to suboptimal performance, especially in cross-dataset scenarios.

Purpose of the Study:

• To develop a robust deepfake detection model that addresses the limitations of existing methods.
• To improve the accuracy and generalizability of deepfake detection, particularly in cross-dataset evaluations.

Main Methods:

• A multi-attention deepfake detection model was proposed, comprising three key components.
• Texture Feature Enhancement: Utilized CondenseNet for efficient texture feature extraction, preserving fine details.
• Multi-Scale Artifact Detection: Incorporated a module to identify manipulated regions at various scales, minimizing identity information impact.
• Multi-Attention Mechanism: Generated multiple attention maps to prioritize image regions and fuse texture and local features for enhanced classification.

Main Results:

• The model demonstrated superior performance on the FaceForensics++ and DFDC benchmarks for facial manipulation detection.
• Achieved state-of-the-art results in cross-dataset evaluations on Celeb-DF-v2, indicating strong generalizability.

Conclusions:

• The proposed multi-attention model effectively integrates local and global features for accurate deepfake detection.
• The method shows significant promise for real-world applications requiring robust detection of manipulated facial content.